Eliza-1

Eliza-1 is the consolidated elizaOS local-inference model repository. Active runtime bundles live under bundles/<tier>/ and are resolved by eliza-1.manifest.json, checksums/SHA256SUMS, evals/aggregate.json, dflash/target-meta.json, and evidence/release.json inside each tier.

The current v1 bundle line packages base/upstream-derived GGUF runtime artifacts and release evidence. Fine-tuned Eliza weights are tracked as the next publishable release state; fine-tuning data, validation reports, and examples live in elizaos/eliza-1-training. Removed legacy long-context experiments are not part of the active release surface.

Active Bundles

Tier	Remote path	Files	Size	Status
0_8b	bundles/0_8b/	72	4.78 GiB	release-candidate; see evidence/release.json and evals/aggregate.json
2b	bundles/2b/	64	7.75 GiB	release-candidate; see evidence/release.json and evals/aggregate.json
4b	bundles/4b/	72	12.78 GiB	release-candidate; see evidence/release.json and evals/aggregate.json
9b	bundles/9b/	85	24.44 GiB	release-candidate; see evidence/release.json and evals/aggregate.json
27b	bundles/27b/	64	41.02 GiB	release-candidate; see evidence/release.json and evals/aggregate.json
27b-256k	bundles/27b-256k/	63	41.02 GiB	release-candidate; see evidence/release.json and evals/aggregate.json

Every active text tier ships both native context and half context variants. Current text contexts are 128k for the default runtime floor and 256k where the tier supports the native context variant.

Runtime Components

Each active bundle is designed as a streaming local-inference pipeline:

• text generation and structured response handling use llama.cpp-compatible GGUF artifacts and Eliza-specific HANDLE_RESPONSE rules;
• MTP / DFlash drafter artifacts live under dflash/ and are validated by dflash/target-meta.json;
• TTS uses omnivoice.cpp-compatible voice artifacts where applicable;
• image generation uses stable-diffusion.cpp-compatible artifacts under imagegen/;
• ASR, VAD, vision, cache, and platform evidence are recorded in each bundle manifest when present.

The structured response contract requires closed action enums, Eliza schema-guided decoding, DFlash prefill support, and deterministic repair evidence. See each tier's evidence/release.json for the exact test reports currently attached to the bundle.

Training And Fine-Tuning

Training data is published in elizaos/eliza-1-training with root JSONL splits, Dataset Viewer parquet mirrors, a native-record validation report, and the smallest-tier fine-tuning runbook at pipeline/docs/training/eliza1-smallest-finetunes.md.

Fine-tuning policy for this release line is size-first: fine-tune only the smallest practical version of each model family first, compare fine-tuned results against the non-fine-tuned baseline, and publish only when the tier's eval gates, backend verification, kernel dispatch reports, checksums, licenses, and Hugging Face upload evidence pass.

Verification Surface

The release audit checks, without downloading large model files:

• required files for each active bundle;
• manifest coverage for text, TTS, ASR, VAD, image generation, vision, DFlash, and cache artifacts;
• native context and half context text entries;
• checksum and LFS hash agreement;
• backend verification for supported CPU, Metal, Vulkan, CUDA, and ROCm surfaces by tier;
• aggregate eval gates for text, voice, ASR, VAD, barge-in, 30-turn, E2E loop, and DFlash evidence;
• MTP / DFlash acceptance evidence;
• structured response evidence;
• dataset schema, privacy attestation, training contract, validation report, and split availability.

Use packages/training/scripts/manifest/audit_hf_eliza1_release.py in the elizaOS repository for the authoritative metadata gate.